Brian Silverman implemented fixed point trig functions for, I forget, LogoWriter or something. But you don't want to get results in integer pixel coordinates, even though that's what the graphics hardware will use. You want units of 1/10,000, or 1/32,768 as that article suggests, because if you round to integer pixels your polygons will end up not closing, because of accumulated error. Instead you have to keep track both of the hardware coordinates and of the more precise pseudo-irrational-valued coordinates.
But you don't want to do that now, when floating point is both fast and 128 bits wide. It's back then, as in all the other cases we're talking about, that breaking the abstraction barrier might have been useful.
I mean, everyone got scared about trusting floating point because of the Pentium divide bug. But that happened because of a last-minute, pretty much literally, maybe last-15-minutes, reversal by an inexperienced engineer of a correction that had already been made to an algorithm, slowing it down a little, but making it correct. The doofus thought, "hey, I can speed this up!" If everyone had followed Intel's very detailed and carefully designed engineering process, it wouldn't have happened.